Solving a parabolic equation

Let us set some global options for all code chunks in this document.

# Set seed for reproducibility
set.seed(1982) 
# Set global options for all code chunks
knitr::opts_chunk$set(
  # Disable messages printed by R code chunks
  message = TRUE,    
  # Disable warnings printed by R code chunks
  warning = TRUE,    
  # Show R code within code chunks in output
  echo = TRUE,        
  # Include both R code and its results in output
  include = TRUE,     
  # Evaluate R code chunks
  eval = TRUE,       
  # Enable caching of R code chunks for faster rendering
  cache = FALSE,      
  # Align figures in the center of the output
  fig.align = "center",
  # Enable retina display for high-resolution figures
  retina = 2,
  # Show errors in the output instead of stopping rendering
  error = TRUE,
  # Do not collapse code and output into a single block
  collapse = FALSE
)
# Start the figure counter
fig_count <- 0
# Define the captioner function
captioner <- function(caption) {
  fig_count <<- fig_count + 1
  paste0("Figure ", fig_count, ": ", caption)
}
# Define the function to truncate a number to two decimal places
truncate_to_two <- function(x) {
  floor(x * 100) / 100
}

my.get.roots <- function(order, beta) {
  mt <- get(paste0("m", order, "table"))
  rb <- rep(0, order + 1)
  rc <- rep(0, order)
  if(order == 1) {
      rc = approx(mt$beta, mt[[paste0("rc")]], beta)$y
    } else {
      rc = sapply(1:order, function(i) {
        approx(mt$beta, mt[[paste0("rc.", i)]], beta)$y
      })
    }
    rb = sapply(1:(order+1), function(i) {
      approx(mt$beta, mt[[paste0("rb.", i)]], xout = beta)$y
    })
    factor = approx(mt$beta, mt$factor, xout = beta)$y
  return(list(rb = rb, rc = rc, factor = factor))
}

m1table <- rSPDE:::m1table
m2table <- rSPDE:::m2table
m3table <- rSPDE:::m3table
m4table <- rSPDE:::m4table

poly_from_roots <- function(roots) {
  coef <- 1
  for (r in roots) {
    coef <- convolve(coef, c(1, -r), type = "open")
  }
  return(rev(coef))
}

compute_sum_poly <- function(factor, pr_roots, pl_roots, cte) {
  
  # Coefficients of p(x) and q(x)
  pr_coef <- poly_from_roots(pr_roots)
  pl_coef <- poly_from_roots(pl_roots)
  
  # Pad to equal length
  max_len <- max(length(pr_coef), length(pl_coef))
  pr_coef <- c(rep(0, max_len - length(pr_coef)), pr_coef)
  pl_coef <- c(rep(0, max_len - length(pl_coef)), pl_coef)
  
  # Compute r(x)
  pr_plus_pl_coef <- factor * pr_coef + cte * pl_coef
  # Roots of r(x)
  pr_plus_pl_roots <- polyroot(rev(pr_plus_pl_coef))

  return(list(new_factor = pr_plus_pl_coef[1], pr_plus_pl_roots = Re(pr_plus_pl_roots)))
}

my.fractional.operators <- function(L, # kappa^2C + G
                                 beta,
                                 C,
                                 scale.factor, # kappa^2
                                 m = 1,
                                 time_step) {

  C <- Matrix::Diagonal(dim(C)[1], rowSums(C)) # lumped
  Ci <- Matrix::Diagonal(dim(C)[1], 1 / rowSums(C)) # lumped 
  I <- Matrix::Diagonal(dim(C)[1])
  L <- L / scale.factor # C + G/kappa^2
  LCi <- L %*% Ci
  roots <- my.get.roots(m, beta)
  Pl.roots <- roots$rb
  Pr.roots <- roots$rc
  factor <- roots$factor
  new_factor_and_roots <- compute_sum_poly(factor, Pr.roots, Pl.roots, time_step)
  
  new_roots <- new_factor_and_roots$pr_plus_pl_roots
  new_factor <- new_factor_and_roots$new_factor
  
  Pr_plus_Pl.factors <- list()
  Pr_plus_Pl.factors[[1]] <- I - LCi * new_roots[1]
  
  if (length(new_roots) > 1) {
    for (i in 2:length(new_roots)) {
      Pr_plus_Pl.factors[[i]] <- I - LCi * new_roots[i]
    }
  }
  Pr.factors <- list()
  Pr.factors[[1]] <- I - LCi * roots$rc[1]

  if (length(roots$rc) > 1) {
    for (i in 2:length(roots$rc)) {
      Pr.factors[[i]] <- I - LCi * roots$rc[i]
    }
  }
  
  #Pr.factors[[length(Pr.factors) + 1]] <- I
   
  output <- list(
    Ci = Ci,
    C = C,
    LCi = LCi,
    L = L,
    m = m,
    beta = beta,
    factor = factor,
    new_factor = new_factor,
    Pr.factors = Pr.factors,
    Pr_plus_Pl.factors = Pr_plus_Pl.factors
  )
  return(output)
}

# install.packages("INLA",repos=c(getOption("repos"),INLA="https://inla.r-inla-download.org/R/testing"), dep=TRUE)
# inla.upgrade(testing = TRUE)
# remotes::install_github("inlabru-org/inlabru", ref = "devel")
# remotes::install_github("davidbolin/rspde", ref = "devel")
# remotes::install_github("davidbolin/metricgraph", ref = "devel")
library(INLA)

## Loading required package: Matrix

## This is INLA_25.05.01 built 2025-05-01 18:43:33 UTC.
##  - See www.r-inla.org/contact-us for how to get help.
##  - List available models/likelihoods/etc with inla.list.models()
##  - Use inla.doc(<NAME>) to access documentation

library(inlabru)

## Loading required package: fmesher

library(rSPDE)

## This is rSPDE 2.5.1
## - See https://davidbolin.github.io/rSPDE for vignettes and manuals.

library(MetricGraph)

## This is MetricGraph 1.4.1
## - See https://davidbolin.github.io/MetricGraph for vignettes and manuals.

## 
## Attaching package: 'MetricGraph'

## The following object is masked from 'package:stats':
## 
##     filter

library(grateful)

library(plotly)

## Loading required package: ggplot2

## 
## Attaching package: 'plotly'

## The following object is masked from 'package:ggplot2':
## 
##     last_plot

## The following object is masked from 'package:stats':
## 
##     filter

## The following object is masked from 'package:graphics':
## 
##     layout

We want to solve the fractional diffusion equation \[\begin{equation} \label{eq:maineq} \partial_t u+(\kappa^2-\Delta_\Gamma)^{\frac{\alpha}{2}} u=f \text { on } \Gamma \times(0, T), \quad u(0)=u_0 \text { on } \Gamma, \end{equation}\] where \(u\) satisfies the Kirchhoff vertex conditions \[\begin{equation} \label{eq:Kcond} \left\{\phi\in C(\Gamma)\;\Big|\; \forall v\in V: \sum_{e\in\mathcal{E}_v}\partial_e \phi(v)=0 \right\} \end{equation}\]

If \(f=0\), then the solution is given by \[\begin{equation} \label{eq:sol_reprentation} u(s,t) = \displaystyle\sum_{j\in\mathbb{N}}e^{-\lambda^{\frac{\alpha}{2}}_jt}\left(u_0, e_j\right)_{L_2(\Gamma)}e_j(s). \end{equation}\]

# Function to build a tadpole graph and create a mesh
gets_graph_tadpole <- function(h){
  edge1 <- rbind(c(0,0),c(1,0))
  theta <- seq(from=-pi,to=pi,length.out = 100)
  edge2 <- cbind(1+1/pi+cos(theta)/pi,sin(theta)/pi)
  edges = list(edge1, edge2)
  graph <- metric_graph$new(edges = edges)
  graph$build_mesh(h = h)
  return(graph)
}

Let \(\Gamma_T = (\mathcal{V},\mathcal{E})\) characterize the tadpole graph with \(\mathcal{V}= \{v_1,v_2\}\) and \(\mathcal{E}= \{e_1,e_2\}\). The left edge \(e_1\) has length 1 and the circular edge \(e_2\) has length 2. As discussed before, a point on \(e_1\) is parameterized via \(s=\left(e_1, t\right)\) for \(t \in[0,1]\) and a point on \(e_2\) via \(s=\left(e_2, t\right)\) for \(t\in[0,2]\). One can verify that \(-\Delta_\Gamma\) has eigenvalues \(0,\left\{(i \pi / 2)^2\right\}_{i \in \mathbb{N}}\) and \(\left\{(i \pi / 2)^2\right\}_{2 i \in \mathbb{N}}\) with corresponding eigenfunctions \(\phi_0\), \(\left\{\phi_i\right\}_{i \in \mathbb{N}}\), and \(\left\{\psi_i\right\}_{2 i \in \mathbb{N}}\) given by \(\phi_0(s)=1 / \sqrt{3}\) and \[\begin{equation*} \phi_i(s)=C_{\phi, i}\begin{cases} -2 \sin (\frac{i\pi}{2}) \cos (\frac{i \pi t}{2}), & s \in e_1, \\ \sin (i \pi t / 2), & s \in e_2, \end{cases}, \quad \psi_i(s)=\frac{\sqrt{3}}{\sqrt{2}} \begin{cases} (-1)^{i / 2} \cos (\frac{i \pi t}{2}), & s \in e_1, \\ \cos (\frac{i \pi t}{2}), & s \in e_2, \end{cases}, \end{equation*}\] where \(C_{\phi, i}=1\) if \(i\) is even and \(C_{\phi, i}=1 / \sqrt{3}\) otherwise. Moreover, these functions form an orthonormal basis for \(L_2(\Gamma_T)\).

# Function to compute the eigenfunctions 
tadpole.eig <- function(k,graph){
x1 <- c(0,graph$get_edge_lengths()[1]*graph$mesh$PtE[graph$mesh$PtE[,1]==1,2]) 
x2 <- c(0,graph$get_edge_lengths()[2]*graph$mesh$PtE[graph$mesh$PtE[,1]==2,2]) 

if(k==0){ 
  f.e1 <- rep(1,length(x1)) 
  f.e2 <- rep(1,length(x2)) 
  f1 = c(f.e1[1],f.e2[1],f.e1[-1], f.e2[-1]) 
  f = list(phi=f1/sqrt(3)) 
  
} else {
  f.e1 <- -2*sin(pi*k*1/2)*cos(pi*k*x1/2) 
  f.e2 <- sin(pi*k*x2/2)                  
  
  f1 = c(f.e1[1],f.e2[1],f.e1[-1], f.e2[-1]) 
  
  if((k %% 2)==1){ 
    f = list(phi=f1/sqrt(3)) 
  } else { 
    f.e1 <- (-1)^{k/2}*cos(pi*k*x1/2)
    f.e2 <- cos(pi*k*x2/2)
    f2 = c(f.e1[1],f.e2[1],f.e1[-1],f.e2[-1]) 
    f <- list(phi=f1,psi=f2/sqrt(3/2))
  }
}

return(f)
}

Implementation of \(u\)

h <- 0.01
graph <- gets_graph_tadpole(h = h)

## Starting graph creation...

## LongLat is set to FALSE

## Creating edges...

## Setting edge weights...

## Computing bounding box...

## Setting up edges

## Merging close vertices

## Total construction time: 0.16 secs

## Creating and updating vertices...

## Storing the initial graph...

## Computing the relative positions of the edges...

T_final <- 2
time_step <- 0.01
time_seq <- seq(0, T_final, by = time_step)
# Compute the FEM matrices
graph$compute_fem()
G <- graph$mesh$G
C <- graph$mesh$C
I <- Matrix::Diagonal(nrow(C))
x <- graph$mesh$V[, 1]
y <- graph$mesh$V[, 2]
edge_number <- graph$mesh$VtE[, 1]
pos <- sum(edge_number == 1)+1
order_to_plot <- function(v)return(c(v[1], v[3:pos], v[2], v[(pos+1):length(v)], v[2]))
weights <- graph$mesh$weights
# Initial condition
# U_0 <- 10*exp(-((x-1)^2 + (y)^2))

kappa <- 1
alpha <- 0.5 # from 0.5 to 2
m = 2
beta <- alpha/2
L <- kappa^2*C + G

my_op <- my.fractional.operators(L, beta, C, scale.factor = kappa^2, m = m, time_step)

my.solver <- function(obj, v){
  Pr.factors <- obj$Pr.factors
  Pr_plus_Pl.factors <- obj$Pr_plus_Pl.factors
  m <- obj$m
  C <- obj$C
  Ci <- obj$Ci
  factor <- obj$factor
  new_factor <- obj$new_factor
  if (m==1){
    temp <- solve(Pr_plus_Pl.factors[[2]], 
                  Pr.factors[[1]] %*% solve(
                    Pr_plus_Pl.factors[[1]], 
                    C%*%v))
  } else if (m==2){
    temp <- solve(Pr_plus_Pl.factors[[3]], 
                  Pr.factors[[2]] %*% solve(
                    Pr_plus_Pl.factors[[2]], 
                    Pr.factors[[1]] %*% solve(
                      Pr_plus_Pl.factors[[1]], 
                      C%*%v)))
  } else if (m==3){
    temp <- solve(Pr_plus_Pl.factors[[4]], 
                  Pr.factors[[3]] %*% solve(
                    Pr_plus_Pl.factors[[3]], 
                    Pr.factors[[2]] %*% solve(
                      Pr_plus_Pl.factors[[2]], 
                      Pr.factors[[1]] %*% solve(
                        Pr_plus_Pl.factors[[1]], 
                        C%*%v))))
  } else if (m==4){
    temp <- solve(Pr_plus_Pl.factors[[5]], 
                  Pr.factors[[4]] %*% solve(
                    Pr_plus_Pl.factors[[4]], 
                    Pr.factors[[3]] %*% solve(
                      Pr_plus_Pl.factors[[3]], 
                      Pr.factors[[2]] %*% solve(
                        Pr_plus_Pl.factors[[2]], 
                        Pr.factors[[1]] %*% solve(
                          Pr_plus_Pl.factors[[1]], 
                          C%*%v)))))
  }
  return((1/factor/new_factor) * Ci %*% temp)
}

op <- fractional.operators(L, beta, C, scale.factor = kappa^2, m = m)
Pl <- op$Pl
Pr <- op$Pr
Ci <- op$Ci
C <- op$C

# Parameters to construct U_0
N_finite <- 4 # choose an even number
adjusted_N_finite <- N_finite + N_finite/2 + 1
EIGENVAL_ALPHA <- NULL
EIGENFUN <- NULL       
for (j in 0:N_finite) {
    lambda_j_alpha <- (kappa^2 + (j*pi/2)^2)^(alpha/2)
    e_j <- tadpole.eig(j,graph)$phi
    EIGENVAL_ALPHA <- c(EIGENVAL_ALPHA, lambda_j_alpha)         
    EIGENFUN <- cbind(EIGENFUN, e_j)            
    if (j>0 && (j %% 2 == 0)) {
      lambda_j_alpha <- (kappa^2 + (j*pi/2)^2)^(alpha/2)
      e_j <- tadpole.eig(j,graph)$psi
      EIGENVAL_ALPHA <- c(EIGENVAL_ALPHA, lambda_j_alpha)         
      EIGENFUN <- cbind(EIGENFUN, e_j)           
    }
}

# Building the initial condition as \sum coeff_j EIGENFUN_j
coeff <- 1*(1:adjusted_N_finite)^-1
coeff[-5] <- 0
# lower_zeroer <- 10  # choose an even number
# adjusted_lower_zeroer <- lower_zeroer + lower_zeroer/2 + 1
# coeff[adjusted_lower_zeroer:adjusted_N_finite] <- 0
U_0 <- EIGENFUN %*% coeff

# Building the true solution as \sum coeff_j EIGENFUN_j e^{-\lambda_j^{\frac{\alpha}{2}}t}
U_true <- matrix(NA, nrow = length(x), ncol = length(time_seq))
for (k in 1:length(time_seq)) {
  aux_k <- rep(0, length(x))
  for (j in 1:adjusted_N_finite) {
    aux_k <- aux_k + exp(-time_seq[k]*EIGENVAL_ALPHA[j])*coeff[j]*EIGENFUN[, j]
  }
  U_true[, k] <- aux_k
}

U_approx <- matrix(NA, nrow = nrow(C), ncol = length(time_seq))
U_approx[, 1] <- U_0

# Time-stepping loop
for (k in 1:(length(time_seq) - 1)) {
  U_approx[, k + 1] <- as.matrix(my.solver(my_op, U_approx[, k]))
}

{r}
Pr.apply.mult <- function(v){return(Pr.mult(op, v))}
Pl.apply.solve <- function(v){return(Pl.solve(op, v))}
PliC <- apply(C, 2, Pl.apply.solve) # PlC^-1
# Precompute the LHS1 matrix
aux <- apply(PliC, 2, Pr.apply.mult) #PrPl^-1C
LHS <- aux + time_step * Matrix::Diagonal(nrow(C)) 

# Initialize U matrix to store solution at each time step
U_approx <- matrix(NA, nrow = nrow(C), ncol = length(time_seq))
U_approx[, 1] <- U_0

# Time-stepping loop
for (k in 1:(length(time_seq) - 1)) {
  # Compute the right-hand side for the second equation
  RHS <- aux %*% U_approx[, k]
  U_approx[, k + 1] <- as.matrix(solve(LHS, RHS))
}

We arrive at the scheme

\[\begin{equation} (P_r^TC+\tau P_\ell^T)U^{k+1} = P_r^TCU^k \label{eq:scheme} \end{equation}\] where \[\begin{equation} P_r = c_m\prod_{i=1}^m (I-r_{1i}C^{-1}L)\quad\text{and}\quad P_\ell = b_{m+1}C\prod_{j=1}^{m+1} (I-r_{2j}C^{-1}L) \end{equation}\] Observe that \[\begin{equation} P_r^T = c_m\prod_{i=m}^1 (I-r_{1i}LC^{-1})\quad\text{and}\quad P_\ell^T = b_{m+1}\prod_{j=m+1}^{1} (I-r_{2j}LC^{-1})\cdot C \end{equation}\] Replacing these two in our scheme we get \[\begin{equation} \left(c_m\prod_{i=m}^1 (I-r_{1i}LC^{-1})+\tau b_{m+1}\prod_{j=m+1}^{1} (I-r_{2j}LC^{-1})\right)CU^{k+1} = c_m\prod_{i=m}^1 (I-r_{1i}LC^{-1})\cdot CU^k \end{equation}\] We can equivalently write this as \[\begin{equation} \left(\dfrac{c_m}{b_{m+1}}\prod_{i=1}^m (I-r_{1i}LC^{-1})+\tau \prod_{j=1}^{m+1} (I-r_{2j}LC^{-1})\right)CU^{k+1} = \dfrac{c_m}{b_{m+1}}\prod_{i=1}^m (I-r_{1i}LC^{-1})\cdot CU^k \end{equation}\]

We need to compute the roots of the polynomial \[\begin{equation} R(x) = \dfrac{c_m}{b_{m+1}} \prod_{i=1}^m (x-r_{1i})+\tau \prod_{j=1}^{m+1} (x-r_{2j}) \end{equation}\] Say \(R\) has leading coefficient \(\tau\) and \(m+1\) roots \(R_k\) for \(k=1,\ldots,m+1\). That is, \[\begin{equation} R(x) = \tau\prod_{k=1}^{m+1} (x-R_k) \end{equation}\] We can then write our scheme as \[\begin{equation} \left(\tau\prod_{k=1}^{m+1} (I-R_kLC^{-1})\right)CU^{k+1} = \dfrac{c_m}{b_{m+1}}\prod_{i=1}^m (I-r_{1i}LC^{-1})\cdot CU^k \end{equation}\] That is, \[\begin{equation} U^{k+1} = \dfrac{c_m}{b_{m+1}}\dfrac{1}{\tau}C^{-1}\prod_{k=1}^{m+1} (I-R_kLC^{-1})^{-1}\prod_{i=1}^m (I-r_{1i}LC^{-1})\cdot CU^k \end{equation}\]

{r}
LHS <- t(Pr)%*% C + time_step * t(Pl)
# Initialize U matrix to store solution at each time step
U_approx <- matrix(NA, nrow = nrow(C), ncol = length(time_seq))
U_approx[, 1] <- U_0

# Time-stepping loop
for (k in 1:(length(time_seq) - 1)) {
  # Compute the right-hand side for the second equation
  RHS <- t(Pr)%*% C %*% U_approx[, k]
  U_approx[, k + 1] <- as.matrix(solve(LHS, RHS))
}

x <- order_to_plot(x)
y <- order_to_plot(y)
max_error_at_each_time <- apply((U_true - U_approx)^2, 2, mean)

U_true <- apply(U_true, 2, order_to_plot)
U_approx <- apply(U_approx, 2, order_to_plot)



# Create interactive plot
fig <- plot_ly()


# Add second line (max_error_at_each_time)
fig <- fig %>% add_trace(
  x = ~time_seq, y = ~max_error_at_each_time, type = 'scatter', mode = 'lines+markers',
  line = list(color = 'red', width = 2),
  marker = list(size = 4),
  name = "Max Error True and Approx 2"
)

# Layout
fig <- fig %>% layout(
  title = "Max Error at Each Time Step",
  xaxis = list(title = "Time"),
  yaxis = list(title = "Max Error"),
  legend = list(x = 0.1, y = 0.9)
)


plot_data <- data.frame(
  x = rep(x, times = ncol(U_true)),
  y = rep(y, times = ncol(U_true)),
  z_true = as.vector(U_true),
  z_approx = as.vector(U_approx),
  frame = rep(time_seq, each = length(x))
)

# Compute axis limits
x_range <- range(x)
y_range <- range(y)
z_range <- range(c(U_true, U_approx))

# Initial plot setup (first frame only)
p <- plot_ly(plot_data, frame = ~frame) %>%
  add_trace(
    x = ~x, y = ~y, z = ~z_true,
    type = "scatter3d", mode = "lines",
    name = "True",
    line = list(color = "blue", width = 2)
  ) %>%
  add_trace(
    x = ~x, y = ~y, z = ~z_approx,
    type = "scatter3d", mode = "lines",
    name = "Approx",
    line = list(color = "red", width = 2)
  ) %>%
  layout(
    scene = list(
      xaxis = list(title = "x", range = x_range),
      yaxis = list(title = "y", range = y_range),
      zaxis = list(title = "Value", range = z_range),
      aspectratio = list(x = 2.4, y = 1.2, z = 1.2),
           camera = list(
      eye = list(x = 1.5, y = 1.5, z = 1),  # Adjust the viewpoint
      center = list(x = 0, y = 0, z = 0))
    ),
    updatemenus = list(
      list(
        type = "buttons", showactive = FALSE,
        buttons = list(
          list(label = "Play", method = "animate",
               args = list(NULL, list(frame = list(duration = 100, redraw = TRUE), fromcurrent = TRUE))),
          list(label = "Pause", method = "animate",
               args = list(NULL, list(mode = "immediate", frame = list(duration = 0), redraw = FALSE)))
        )
      )
    ),
    title = "Time: 0"
  )

# Convert to plotly object with frame info
pb <- plotly_build(p)

# Inject custom titles into each frame
for (i in seq_along(pb$x$frames)) {
  t <- time_seq[i]
  err <- signif(max_error_at_each_time[i], 4)
  pb$x$frames[[i]]$layout <- list(title = paste0("Time: ", t, " | Max Error: ", err))
}

fig  # Display the plot

pb

Figure 1: Caption